Product decisions7 min readOctober 12, 2026

Amazon Nova Pro vs Nova Lite vs Nova 2 Lite for reading business documents

Amazon's own Nova models are the cheapest way to ask questions about documents on Amazon Bedrock. Here is how Nova Pro, Nova Lite and the newer Nova 2 Lite differ on context, output length, routing, lifecycle, caching and price, and which one fits which document question.

JeVaughn Ferguson
Founder, developer

The short version

Nova Lite is the cheapest way to answer factual questions about a document on Bedrock, Nova Pro costs about thirteen times more and handles judgement better, and Nova 2 Lite adds a 1M-token window and 64K-token answers but has no in-Region option. All three cache at most 20,000 tokens for five minutes, so long documents are re-read in full on every question. Nova Pro and Nova Lite are past their end-of-life floor while still active, so build for a model swap, and pick the model by where your data may be processed as much as by price.

Searches for "amazon nova pro" and "amazon nova lite" are both up about 150% over the past year, and much of that interest comes from teams pricing out document questions on Amazon Bedrock. Nova is Amazon's own model family, sold on Bedrock next to Claude, OpenAI and others, and for reading a contract or an invoice it is usually the cheapest option on the list.

Three Nova models matter for documents: Nova Pro, Nova Lite and Nova 2 Lite. They share a name and a caching scheme, but they differ in ways that decide where your data is processed and how long you can rely on them. Everything below comes from the Bedrock model cards as read on 2026-10-12.

Which Nova models read documents

The Nova name covers more than text models. Nova Sonic is a speech model, Nova Act drives a web browser, and Nova Micro is text-only, so none of those take a PDF. The three that accept text, images, video and the Converse document block are Nova Lite, Nova Pro and Nova 2 Lite.

Nova Lite and Nova Pro launched together on 5 December 2024 with a knowledge cutoff of October 2024. Nova 2 Lite launched on 2 December 2025 with a cutoff of October 2025. All three are on the bedrock-runtime endpoint and work with the same Converse call, so switching between them is a change of model ID.

Nova Lite    amazon.nova-lite-v1:0     geo: us. / ca. / eu. / apac.
Nova Pro     amazon.nova-pro-v1:0      geo: us. / eu. / apac.
Nova 2 Lite  amazon.nova-2-lite-v1:0   geo: us. / eu. / jp.   global: global.

Which Amazon Bedrock model should answer questions about your documents?

Context and answer length

Nova Lite and Nova Pro each have a 300K-token context window and a 10K-token output limit. Nova 2 Lite has a 1M-token window and up to 64K output tokens.

For one document at a time the difference rarely matters. The Converse document block caps each file at 4.5 MB, and a 30-page report is around 11,000 tokens, so 300K is already far more than a single file needs. Nova 2 Lite's larger window starts to count when you send several documents and a long conversation together, and its larger output limit helps if you ask for a full rewrite or a long structured extract rather than an answer.

Nova Lite and Nova Pro: 300K context, 10K output.
Nova 2 Lite: 1M context, 64K output.
One 4.5 MB document: well inside every window.

Where the request is processed

This is the difference most teams miss. Nova Pro and Nova Lite can be called in-Region in some Regions, so the request stays where you sent it. Nova Pro has an in-Region option in US East (N. Virginia), London, Sydney, Jakarta, UAE and GovCloud; Nova Lite in those plus Ohio, Oregon, Stockholm and Tokyo. Neither supports global inference.

Nova 2 Lite has no in-Region option in any Region on its model card. It is served only through geo profiles, which keep requests inside a geography such as the US or EU, or through a global profile that can route anywhere. If a client contract says documents must be processed in one Region, Nova 2 Lite cannot meet it today, and Nova Pro or Nova Lite called in-Region can.

Even the geo profiles move data between Regions. The US profile for Nova Pro, called from N. Virginia, can route to N. Virginia, Ohio or Oregon. That is usually fine for US residency and not fine for single-Region commitments.

Amazon Bedrock data retention modes for business documents

Single-Region requirement: Nova Pro or Nova Lite with the in-Region model ID.
US or EU residency: any of the three through its geo profile.
No residency requirement: Nova 2 Lite's global profile is an option.

Lifecycle: the older two are past their floor

Every Bedrock model card states an end-of-life date that AWS will not go below. For Nova Pro and Nova Lite that floor was 5 December 2025, and it has passed. Both are still listed as Active, with a legacy period of at least six months once retirement is announced, so nothing breaks today. But they are the models most likely to get a retirement notice next.

Nova 2 Lite's floor is 2 December 2026. In practice that means whichever Nova you choose, you should expect to swap models within a year and keep your prompt and document handling model-independent. With Converse, the swap is a configuration change; what needs re-testing is answer quality on your own documents.

Nova Pro and Nova Lite: EOL no sooner than 5 Dec 2025 (passed, still active).
Nova 2 Lite: EOL no sooner than 2 Dec 2026.
Watch the Bedrock model lifecycle page for retirement notices.

Caching is the same, and it is small

All three Nova models share one prompt caching scheme on Bedrock: a 1,000-token minimum per checkpoint, up to four checkpoints, a 5-minute TTL, and a 20,000-token maximum for the cached prefix. Nova also caches text prompts implicitly without any setting.

The 20,000-token ceiling is the constraint for documents. A document of roughly 50 text pages or more will not fit, so every follow-up question pays full price for the whole file again. For short contracts and invoices, caching works and a follow-up costs a fraction of the first question.

What a document chat costs on Amazon Bedrock, by model, with prompt caching

Features and service tiers

Nova Pro has the widest Bedrock feature list of the three: Guardrails, Agents, Flows, prompt management, model evaluation and intelligent prompt routing. Nova Lite's card lists Guardrails, streaming and client-side tool calling. Nova 2 Lite lists Guardrails, abuse detection, prompt optimization and tool calling, but not Agents, Flows or Knowledge Bases.

Service tiers differ too. Nova Pro and Nova 2 Lite offer Standard, Priority and Flex; Nova Lite offers Standard only. Flex is cheaper for work that can wait, such as a nightly batch that summarises the day's uploads.

Nova Pro: Standard, Priority, Flex; Agents and Flows supported.
Nova Lite: Standard only.
Nova 2 Lite: Standard, Priority, Flex; no Agents, Flows or Knowledge Bases on its card.

What a question costs

At the US list prices we last checked, Nova Lite costs $0.06 per million input tokens and $0.24 per million output tokens, and Nova Pro costs $0.80 and $3.20. A question about a 20-page contract, about 15,000 input tokens with an 800-token answer, comes to about $0.0011 on Nova Lite and about $0.015 on Nova Pro. Nova Pro is roughly thirteen times the price, and still under two cents per question.

We have not quoted Nova 2 Lite's price because the Bedrock pricing page renders its tables in a way we could not verify; check the page before you budget it.

20-page contract, ~15,000 tokens in, 800 out
Nova Lite: 15,000 x $0.06/M + 800 x $0.24/M = ~$0.0011
Nova Pro:  15,000 x $0.80/M + 800 x $3.20/M = ~$0.015

Amazon Bedrock pricing for document Q&A Keep document AI spending under control in your own AWS account

Which Nova for which question

Use Nova Lite for look-ups: the invoice total, the renewal date, the name of the supplier. It is fast, very cheap and accurate enough when the answer is a fact on the page. Use Nova Pro when the question needs judgement or comparison: whether a clause differs from your standard, what the risks in a lease are, how two versions of a policy differ. Consider Nova 2 Lite when you need a very long output or several large files in one conversation and you have no single-Region requirement.

Look-ups on short documents: Nova Lite.
Judgement, comparison, summaries for a manager: Nova Pro.
Long outputs or several big files, no residency constraint: Nova 2 Lite.
Documents over ~50 pages with many follow-ups: consider Claude, whose cache holds them.

How BucketDesk Document AI uses Nova, and how it compares with Amazon Q

BucketDesk Document AI uses Amazon Nova Pro by default, through the US geo inference profile, and lets you switch a connection to Nova Lite or another Bedrock model such as Claude Haiku 4.5 in its settings. It runs on Amazon Bedrock in your own AWS account through a separate document chat role, so AWS bills the tokens to you at list price and your account's data retention mode applies. You open a file in a connected bucket, ask in plain language, and get an answer with a citation that highlights the supporting text. It is part of the Ultra plan at $350 a month with 10 operators and unlimited viewers.

The AWS assistants take a different shape. Amazon Q Business is closed to new customers, and its replacement, Amazon Quick Suite, indexes a bucket on a schedule and answers across the whole knowledge base, and the S3 integration needs the Enterprise plan. That suits questions across thousands of files. For a question about the file in front of you, picking the model yourself and paying per question is usually cheaper and more predictable.

Amazon Quick Suite for documents in S3: pricing and setup Document AI on the features page

Amazon Quick Suite: indexed, cross-file answers, per-user plus per-account pricing.
Amazon Q Business: not available to new customers.
BucketDesk Document AI: one file at a time with citations, Nova Pro by default or a model you choose, Bedrock tokens on your own AWS bill.

Try it in BucketDesk

Minimum is free. Deploy a scoped role with CloudFormation, sign in, and browse, without handing anyone an access key.

Connect a bucket

Primary sources

AWS FinOps Agent finds the cost. Knowing what to fix, and fixing it, is the hard part

The FinOps Agent tells you why the bill moved and lists what AWS recommends. Someone still has to choose the right change, make it without breaking anything, and check next month that the money actually came off.FinOps Agent preview scope (Cost Anomaly Detection, CloudTrail correlation, Cost Optimization Hub and Compute Optimizer summaries, Jira and Slack delivery), the hand-off point, and how BucketDesk's S3 Savings tool covers assessment, one-click correction and verified outcomes with timelines.

Discussion

0 comments · open to guests · moderated

Liked this? Get the next article by email. No schedule, no filler, one click to leave.